The Understanding Layer of AI

Published May 26, 2026, 8:43 a.m.

AI can summarize research, write code, and draft business plans... but it doesn't understand what it's doing. And for many people, it's already becoming a crutch.

AI still requires a human to think critically about its output and ask, "does this make sense?" We have all heard stories of people who didn't fact check or even look at the results. If you only rely on AI to think for you, then you have become replaceable. You are the understanding layer of AI.

AI can help you to deepen your understanding while you use it. Below are some thoughts on how to do that.

Ask for More Solutions

Often I will ask AI to give me several possibilities, with pros and cons. Typically in life there isn’t just one right solution. You want to think through what is the best way. AI can help you think through topics. Telling it over and over to give you different solutions also helps you discover the various possibilities. Another trick is to ask, "What am I missing?"

I use this concept when I want to code an algorithm for trading. There are many design possibilities for how to code an algorithm. Thus, I will ask for different approaches with the various trade offs. Sometimes I will even push back on a proposed method. The back and forth helps me to work through the design choices and surfaces options I would not have thought of on my own.

Review the Diff

Don’t have it fix things for you. Go through each change, understand it, and accept/decline it deliberately. Yes, this takes longer, but I regularly catch AI changing things I didn't ask it to. There are a lot of ways to do this:

Make sure to prompt it to not just show you the change, but why AI thinks a change is needed.

Deterministic Outputs

If a question/prompt has a precise, deterministic answer, don't ask AI to answer it. Examples, might be a calculation, lookup, transformation... Anything with a definitive answer. Ask AI to create code that answers the question. Done this way, you can validate the logic of the code, test it, and trust it forever. Whereas AI answers will always need continual validation.

Last week I needed to calculate rolling Sharpe ratios across 15 years of futures data. I could have pasted the data into Claude and asked for the answer. I would likely get something plausible but unverifiable. Instead, I had AI write a simple Python script using Polars. I read the code, checked the formula, and ran it. Now I have a function I can reuse, audit, and trust.

Compare AI Models on Complex Questions

When I want to think deeply about something, I run the same prompt through multiple AI models. The differences can be enlightening and uncover insights you would not have understood from using just one. Business plans are a great use case, each model brings different blind spots. Seeing the insights side by side surfaces what you might otherwise miss.

Devil's Advocate

After a prompt ask AI to tell you the strongest case against its own answer. AI is often sycophantic, so this helps to protect against that. Remember, you want the right answers, not to feed your ego.

The Understanding Layer

One example from my own experience, I have been teaching myself machine learning. I can write Python so doing machine learning is relatively easy. What I realized is without understanding what the algorithms are actually doing, it quickly becomes garbage in, garbage out. Just because I can run a Support Vector Machine, does not mean that I can apply it effectively. Thus, I took a step back to understand the math behind what is happening. Has it taken longer? Yes, I fell down a rabbit hole of learning math all over again. But in the long run, I will more effectively apply machine learning. I can validate results.

Using AI this way won't make you faster. But your output will be deeper, more thoughtful, and more defensible. The people I see using AI most effectively all say some version of this. Thinking of AI as only a productivity tool limits the possibilities of how it can truly help you.

If you find yourself in your job passing AI output along, AI will eventually do your job. It already is. The value you add is the understanding layer... the judgment, the context, the validation, the "does this actually make sense for our situation?" That's the part that can't be automated, and it's the part we should be deliberately strengthening.

Similar Posts

First Post

0 comments

Be the first to comment!

Add a new comment